LogFAQs > #979726611

[LFAQs-redacted-quote]


The issue here was that there is a massive, massive difference between building an object recognition model that can tell you that a banana is a banana when it's in clean lighting on a table at a flat angle and one that can identify a banana at any arbitrary distance or angle while also likely being partially occluded due to bodies being in the way. And then managing to make that perfectly accurate model work for every product in the store.

From what I remember, they used other things like weight plates to supposedly help the algorithm, but what they were trying to do was never going to be 100% solvable just using computer vision anyway. Stealing anything from the store would literally just be as easy as standing in between the camera and the object and then putting it in your cart if they tried that.